Abstract

The classical matched filter,¹ originally derived to be optimal for detecting a signal in noise, can be far from satisfactory in image recognition work if the problem is to detect a given signal, i.e., object, among many other signals, i.e., competing objects. In contexts ranging from computer vision in robot assembly (nut or bolt?) to military detection and targeting (tent or tank?), a persistent complication has been susceptibility to false indications, with consequences potentially ranging from the ineffectual (stopping an assembly line) to the tragic (bombing the wrong target). Even if the alternative objects likely to cause false indications are known a priori, the design of a filter to minimize incorrect discrimination entails substantially greater complexity²; and if the alternative objects are unknown, the problem's difficulty is exacerbated further. Therefore, a new filtering strategy, based on emphasizing those object characteristics which contribute to distinctiveness while suppressing irrelevant features, was developed. The filter design requires no knowledge of competing alternative objects and hence maintains computational feasibility even for quite complicated images.

© 1987 Optical Society of America